def learning():
# add to learning dataset
files = []
for i in [1, 3, 5]:
for j in range(1, 8):
filename = r'.\image\bdd\pose' + str(j) + '_' + str(i) + '.png'
files.append([filename, j])
data = []
for file in files:
print(file[0])
nbClass = file[1]
img = lireFichier(file[0])
graph = skeleton(img)
firstFilter(graph)
secondFilter(graph)
thirdFilter(graph)
deleteCircle(graph)
data_head, data_lefthand, data_righthand, data_leftleg, data_rightleg = calculateDataset(
graph)
percentage_length_left = data_lefthand[3] / data_lefthand[2]
percentage_angle_left = data_lefthand[4] / 90
percentage_length_right = data_righthand[3] / data_righthand[2]
percentage_angle_right = data_righthand[4] / 90
data.append([
percentage_length_left, percentage_angle_left,
percentage_length_right, percentage_angle_right, nbClass
])
# data.append([data_head, data_lefthand, data_righthand, data_leftleg, data_rightleg])
# drawGraph(graph)
return data
def main():
# Load data
import cv2
import glob
img_dir = "train_custom/" # Enter Directory of all images
data_path = os.path.join(img_dir, '*g')
files = glob.glob(data_path)
data = []
for f1 in files:
img = rgb2gray(cv2.imread(f1))
data.append(img)
# Preprocessing
print("Start to data preprocessing...")
data = [preprocessing(d) for d in data]
# Create Hopfield Network Model
model = network.HopfieldNetwork()
model.train_weights(data)
# Generate testset
img_dir = "test_custom/" # Enter Directory of all images
data_path = os.path.join(img_dir, '*g')
files = glob.glob(data_path)
test = []
# Since same name order will be the same
for f1 in files:
img = rgb2gray(cv2.imread(f1))
test.append(img)
test = [preprocessing(d) for d in test]
predicted = model.predict(test, threshold=0, asyn=False)
print("Show prediction results...")
plot(data, test, predicted)
def get_train_data(sample_path, vector_stub):
os.chdir(sample_path)
data = []
label = []
target_set = set(
[elem for elem in os.listdir(sample_path) if os.path.isdir(elem)])
print target_set
class_num = len(target_set)
class_dict = dict(zip(target_set, range(class_num)))
for root, dirs, files in os.walk(sample_path, topdown=False):
word_data = [
get_icf_feature(cv2.imread(os.path.join(root, name)), vector_stub)
for name in files if os.path.splitext(name)[-1] in EXT_DICT
]
for elem in word_data:
label.append(class_dict[root.split(os.sep)[-1]])
data.append(elem)
data = np.array(data)
label = np.array(label)
return data, label, class_dict
def load_parallel(files, id_label, path, test_model):
process_list = list()
X = []
Y = []
files = cv.n_list(files, n_cores)
for id_core in range(n_cores):
p = multiprocessing.Process(target=load_X_compress_parallel,
args=(files[id_core], id_label, path,
test_model, id_core))
p.start()
process_list.append(p)
# Join all the threads
for pp in process_list:
pp.join()
for i in range(n_cores):
X_core, Y_core = (return_process_dict[i])
X.append(X_core)
if test_model == True:
Y.append(Y_core)
# resultado dos cores na mesma ordem do ids de entradas
data = []
labels = []
for core in range(n_cores):
for index in range(len(X[core])):
data.append(X[core][index])
if test_model == True:
labels.append(Y[core][index])
return data, labels
def main():
#This code sets up the parser for command line arguments specifying parameters for creating the h5py file.
parser = argparse.ArgumentParser()
parser.add_argument("data_f")
parser.add_argument("label_f")
parser.add_argument("-o",
"--output",
action='store',
default="data.h5",
help='output data dir and filename')
parser.add_argument("-f",
"--force",
action='store_false',
help="overwrite output file")
args = parser.parse_args()
OUTPUT_FILE = args.output
DATA_DIR = args.data_f
LABEL_FILE = args.label_f
SOURCE_IMAGES = os.path.join(DATA_DIR)
images = glob(os.path.join(SOURCE_IMAGES, "*.jpg"))
labels = pd.read_csv(LABEL_FILE, na_values='-')
#np.savez("asdfasdf",labels.to_numpy())
labels.to_hdf("labeldata.h5", "labels", mode="w")
#reread = pd.read_hdf('./labeldata.h5')
data = []
print("d:{} lb:{} of:{} img:{}".format(args.data_f, args.label_f,
OUTPUT_FILE, images))
print("{}".format(labels.to_numpy()[0]))
#if path.isfile(OUTPUT_FILE) and args.force:
# Image information and resizing
NUM_IMAGES = len(images)
HEIGHT = 3024 // 4
WIDTH = 4032 // 4
CHANNELS = 3
SHAPE = (HEIGHT, WIDTH, CHANNELS)
with h5py.File('imagedata.h5', 'w') as hf:
for i, img in enumerate(tqdm(images)):
image = io.imread(img)
image = resize(image, (WIDTH, HEIGHT), anti_aliasing=True)
data.append(image)
np.savez("progressivesave", images=data[i])
#data.to_hdf("imagedata.h5", "images",mode="w")
#want to save a new image to h5 one at a time like the npz
# imgs = hf.create_dataset(
# name='images',
# data=data,
# #shape=(HEIGHT, WIDTH, CHANNELS),
# #maxshape=(HEIGHT, WIDTH, CHANNELS),
# compression="gzip",
# compression_opts=9)
hf.close()
imgs = np.array(data)
np.savez("dataimages", images=data)
def fillNetData(idClass, elements, absClassPath, absdataDirPath):
'''
idClass: class id to identify each img in elements
elements: list of images
absClassPath: absolute path to the current images holder (class)
absdataDirPath: target path to copy all images in elements
retunrs a list of string pairs: absoluteImagePath idClass
'''
data = []
for img in elements:
# absolute path to current the image
absFilePath = os.path.join(absClassPath, img)
# path to caffe image holder
targetPath = os.path.join(absdataDirPath, img) + '.jpg'
# Save the image in the caffe image holder
resizeAndSave(absFilePath, SIZE, SIZE, targetPath)
# to copy the pic without RESIZE uncomment next line and comment the previous line
#shutil.copy(absFilePath, absdataDirPath)
# add the image path and their class to train
data.append(str(targetPath) + ' ' + str(idClass))
return data
def addArtificialData():
print "here"
baseName = os.path.basename(leftEyePath)
print baseName
data_dir = os.path.join(projectPath,baseName)
print data_dir
files = os.listdir(data_dir)
files = [f for f in files if f.split('.')[-1]=='txt']
print files
data = []
for f in files:
label = f.split('.')[0]
filePath = os.path.join(data_dir,f)
with open(filePath,'r') as r:
for image in r:
data.append(image.strip())
#print data
for f in data:
parentDir = os.path.dirname(f)
image_name = f.split('/')[-1].split('.')[0]
scale_image_name = os.path.join(parentDir,image_name+'_s.jpg')
roate_image_name = os.path.join(parentDir,image_name+'_r.jpg')
print image_name
img = io.imread(f,as_grey=True)
scale_image = rescale(img,0.9)
rotated_image = rotate(img,5,resize=False)
print img.shape
print scale_image.shape
print rotated_image.shape
io.imsave(scale_image_name,scale_image)
io.imsave(roate_image_name,rotated_image)
raw_input()
def segment_data(self, frame, segments):
data = []
self.superpixel.fill(0)
for i in range(1 + np.max(segments)):
pts = np.where(segments == i)
col = np.average(frame[pts[0], pts[1]], axis=0).astype(np.uint8)
#center of contour
self.tmp.fill(0)
self.tmp[pts] = 255
c = cv2.findContours(self.tmp, cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)[-2][0]
M = cv2.moments(c)
if abs(M["m00"]) > 1e-3:
cX = int(M["m10"] / M["m00"])
cY = int(M["m01"] / M["m00"])
self.superpixel[pts] = col
data.append({
'idx': i,
'size': np.count_nonzero(pts),
'color': col,
'center': (cY, cX)
})
return data
def draw_pic(rgb):
data = []
for i in range(rgb.shape[0]):
for j in range(rgb.shape[1]):
data.append(int(rgb[i][j]))
plt.hist(data, bins=256, facecolor="blue", edgecolor="black", alpha=0.7)
plt.show()
def populate_3d_graph(
dummy2_children,
show_hide_seg_3d,
drawn_shapes_data,
last_3d_scene,
last_render_id,
image_display_top_figure,
image_display_side_figure,
):
# extract which graph shown and the current render id
graph_shown, current_render_id = dummy2_children.split(",")
current_render_id = int(current_render_id)
start_time = time.time()
cbcontext = [p["prop_id"] for p in dash.callback_context.triggered][0]
# check that we're not toggling the display of the 3D annotation
if cbcontext != "show-hide-seg-3d.children":
PRINT("might render 3D, current_id: %d, last_id: %d" %
(current_render_id, last_render_id))
if graph_shown != "3d shown" or current_render_id == last_render_id:
if current_render_id == last_render_id:
PRINT("not rendering 3D because it is up to date")
return dash.no_update
PRINT("rendering 3D")
segs_ndarray = shapes_to_segs(
drawn_shapes_data,
image_display_top_figure,
image_display_side_figure,
).transpose((1, 2, 0))
# image, color
images = [
(img.transpose((1, 2, 0))[:, :, ::-1], "grey"),
]
if show_hide_seg_3d == "show":
images.append((segs_ndarray[:, :, ::-1], "purple"))
data = []
for im, color in images:
im = image_utils.combine_last_dim(im)
try:
verts, faces, normals, values = measure.marching_cubes(im,
0,
step_size=3)
x, y, z = verts.T
i, j, k = faces.T
data.append(
go.Mesh3d(x=x,
y=y,
z=z,
color=color,
opacity=0.5,
i=i,
j=j,
k=k))
except RuntimeError:
continue
fig = go.Figure(data=data)
fig.update_layout(**last_3d_scene)
end_time = time.time()
PRINT("serverside 3D generation took: %f seconds" %
(end_time - start_time, ))
return (fig, current_render_id)
def HOG_data_measurement(file_list_):
data = []
for path in file_list:
grey_img = load_image_and_pre_processing(path)
hog_data,hog_image = hog(grey_img, orientations=8, pixels_per_cell=(8, 8),
cells_per_block=(1, 1), visualise=True)
data.append(hog_data)
return data
def get_HOG_Features(trainingPath, testingPath, cell_size=16, bin_size=8):
from hog import Hog_descriptor
# initialize the local binary patterns descriptor along with the data and label lists
data = []
labels = []
test_data = []
test_labels = []
start_time = time.time()
# loop over the training images
for imagePath in paths.list_files(trainingPath, validExts=(".png",".ppm")):
# open image
img = cv2.imread(imagePath)
gray = np.matrix(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY))
resized_image = cv2.resize(gray, (48, 48))
# get hog features
hog = Hog_descriptor(resized_image, cell_size=cell_size, bin_size=bin_size)
vector = hog.extract()
v = np.array(vector)
# extract the label from the image path, then update the
# label and data lists
labels.append(int(imagePath.split("/")[-2]))
data.append(vector)
# loop over the testing images
for imagePath in paths.list_files(testingPath, validExts=(".png",".ppm")):
# open image
img = cv2.imread(imagePath)
gray = np.matrix(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY))
resized_image = cv2.resize(gray, (48, 48))
# get hog features
hog = Hog_descriptor(resized_image, cell_size=cell_size, bin_size=bin_size)
vector = hog.extract()
# extract the label from the image path, then update the
# label and data lists
test_labels.append(int(imagePath.split("/")[-2]))
test_data.append(vector)
feature_extraction_runtime = (time.time() - start_time)
data = np.array(data)
labels = np.array(labels)
test_data = np.array(test_data)
test_labels = np.array(test_labels)
print "[INFO] HOG Features are ready!"
print "Total image: ", len(data) + len(test_data)
print "Feature extraction runtime: ", feature_extraction_runtime
print "Average for one image:", feature_extraction_runtime / (len(data) + len(test_data))
return (data, labels, test_data, test_labels)
def get_LBP_Features(trainingPath, testingPath ,p=24, r=8):
from localbinarypatterns import LocalBinaryPatterns
from sklearn.utils import shuffle
# initialize the local binary patterns descriptor along with the data and label lists
desc = LocalBinaryPatterns(p, r)
data = []
labels = []
test_data = []
test_labels = []
start_time = time.time()
# loop over the training images
for imagePath in paths.list_files(trainingPath, validExts=(".png",".ppm")):
# load the image, convert it to grayscale, and describe it
image = cv2.imread(imagePath)
gray = np.matrix(cv2.cvtColor(image, cv2.COLOR_BGR2GRAY))
resized_image = cv2.resize(gray, (32, 32))
hist = desc.describe(resized_image)
hist = hist / max(hist)
# extract the label from the image path, then update the
# label and data lists
labels.append(int(imagePath.split("/")[-2]))
data.append(hist)
# loop over the testing images
for imagePath in paths.list_files(testingPath, validExts=(".png",".ppm")):
# load the image, convert it to grayscale, describe it, and classify it
image = cv2.imread(imagePath)
gray = np.matrix(cv2.cvtColor(image, cv2.COLOR_BGR2GRAY))
resized_image = cv2.resize(gray, (32, 32))
hist = desc.describe(resized_image)
hist = hist / max(hist)
# extract the label from the image path, then update the
# label and data lists
test_labels.append(int(imagePath.split("/")[-2]))
test_data.append(hist)
feature_extraction_runtime = (time.time() - start_time)
data = np.array(data)
labels = np.array(labels)
#test_data = np.array(test_data)
#test_labels = np.array(test_labels)
data, labels = shuffle(data,labels)
print "[INFO] LBP Features are ready!"
print "Total image:", len(data) + len(test_data)
print "Feature extraction runtime:", feature_extraction_runtime
print "Average for one image:", feature_extraction_runtime / (len(data) + len(test_data))
return (data, labels, test_data, test_labels)
def doPreProcessing(images, path):
data = []
sizeOfData = len(images)
for i, image in enumerate(images):
img = cv2.imread(path + image, cv2.IMREAD_GRAYSCALE)
img = cv2.resize(img, (IMG_SIZE, IMG_SIZE), interpolation= cv2.INTER_LINEAR)
imgData, _ = hog(image, orientations=8, pixels_per_cell=(10, 10),cells_per_block=(2, 2),visualize=True,block_norm='L2')
data.append([val[0] for val in imgData.tolist()])
updateProgress(i, sizeOfData)
return np.array(data)
def train_many_svm(hog, data, labels):
data = []
for i in dataImg:
data.append(calculate_hog(hog, i))
train_data, eval_data, train_labels, eval_labels = train_test_split(data, labels, test_size=0.2, shuffle=True)
scores = []
cList = (0.1, 0.5, 1, 2, 4, 6, 8, 10, 100, 200)
X = np.arange(0.05, 2, 0.1)
tolList = (1e-4, 1e-3, 1e-2, 0.1, 1, 10)
Y = np.arange(0.05, 2, 0.1)
max_score = 0
max_params = ""
print ("Number of training samples: ", len(train_data), "\nNumber of samples for evaluation: ", len(eval_data))
for param1 in X.tolist():
for param2 in Y.tolist():
start = time.time()
clf = svm.SVC(C=param1, tol=1.55, cache_size=100, kernel='rbf')
data_scaler = StandardScaler(copy=True, with_mean=True, with_std=True)
data_scalerTransf = data_scaler.fit_transform(train_data)
clf.fit(data_scalerTransf,train_labels)
data_scalerTransf = data_scaler.fit_transform(train_data)
numOk = 0
eval_data_transform = []
for j in range (len(eval_data)):
trans_data = data_scaler.transform(eval_data[j].reshape(1,-1))
eval_data_transform.append(trans_data)
prediction = clf.predict(trans_data)
if (prediction == eval_labels[j]): numOk += 1
score2 = clf.score(eval_data_transform, eval_labels)
finish = time.time()
score = np.around(numOk/len(eval_data)*100,2)
if score > max_score:
max_params = 'Score: '+ str(score), " %, param1 = ", str(param1), "param2 = ", str(param2)
max_score = score
scores.append(score)
print ('Score: ', score, " %, param1 = ", param1, "param2 = ", param2)
print ('Time: ', finish-start)
print (max_params)
X, Y = np.meshgrid(X,Y)
Z = np.asarray(scores).reshape(X.shape[0], Y.shape[1])
fig = plt.figure()
ax = Axes3D(fig)
ax.set_xlabel('C')
ax.set_ylabel('tol')
ax.set_zlabel('result')
ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=cm.viridis)
plt.show()
return clf, data_scaler
def preprocess_frames(frames,numbers,process_list,ht,wd,channels,ROI_array=None):
data=[]
for x,i in zip(frames, range(0,frames.__len__())):
#print(x,i)
img=None
#FIrst loading image
if channels==1:
img=cv2.imread(x,0) #Use this for RGB to GS #TODO change this based on dset
img=img.reshape(img.shape[0],img.shape[1],1)
# print(img.shape
elif channels==3:
# print("Thermal pose IUV")
img=cv2.imread(x,1)#1 for color images
#resize
# img=img.reshape(ht,wd,3)
if 'ROI_frame' in process_list:
box=ROI_array[i,:]
left, top, right, bottom=int(box[1]),int(box[0]),int(box[3]),int(box[2])
#Pixel outside the ROI are assigned -1, GAN inpu range []-1,1]
out=-np.ones(shape=img.shape,dtype='float32')
patch=img[top:bottom,left:right,:]
#Normalizing only inside ROI
if 'Processed' in process_list:
mean=np.mean(patch,axis=(0,1))
patch=patch-mean
patch=patch/ 255.0
out[top:bottom,left:right,:]=patch
img=cv2.resize(out,(ht,wd))
img=img.reshape(ht,wd,channels)#resize on may remove the channels dis
elif 'Processed' in process_list:
img=cv2.resize(img,(ht,wd))
img=img.reshape(ht,wd,channels)
img=img-np.mean(img,axis=(0,1))#Mean centering
img=img.astype('float32') / 255. #normalize
else:
img=cv2.resize(img,(ht,wd))
img=img.reshape(ht,wd,channels)
data.append(img)
#data = data.reshape((len(data), np.prod(data.shape[1:]))) #Flatten the images
data=np.array(data)
print('data.shape', data.shape)
# print(numbers)
return data,numbers,frames
def load_set(folder, shuffle=False):
img_list = sorted(glob.glob(os.path.join(folder, '*.png')) + \
glob.glob(os.path.join(folder, '*.jpg')) + \
glob.glob(os.path.join(folder, '*.jpeg')))
if shuffle:
np.random.shuffle(img_list)
data = []
filenames = []
for img_fn in img_list:
img = load_image(img_fn)
data.append(img)
filenames.append(img_fn)
return data, filenames
def generate_(data_range):
'''
generate a data in form [Fx,Fy], chosen randomly from @data_range
'''
data = []
x = np.random.choice(data_range, 1)
x = x[0]
y = np.random.choice(data_range, 1)
y = y[0]
# y=-x
data.append(float(x))
data.append(float(y))
return data
def loadData(parentDir):
'''From the parentDir get the dir name where .txt files are stored.
This function will return data and the labels associated with it'''
baseName = os.path.basename(parentDir)
data_dir = os.path.join(projectPath,baseName)
files = os.listdir(data_dir)
files = [f for f in files if f.split('.')[-1]=='txt']
data = []
for f in files:
label = f.split('.')[0]
filePath = os.path.join(data_dir,f)
with open(filePath,'r') as r:
for image in r:
data.append([image.strip(),label])
return data
def extract_info(self, img, index, perc):
data = []
#print(img[0])
for i in range(0, index):
data.append(img[i//self.x][i%self.x])
b = index//8
with open('lsb_data/lsb_perc{}_data.txt'.format(perc), "wb") as f:
for i in range(0, b):
tmp = data[i*8 : i*8+8]
t = 0
n = 7
for j in tmp:
t += (j%2) * 2**n
n -= 1
f.write(bytes([t]))
def dataImage(request):
data = []
for x in range(1, 451):
data.append("/static/data/a (" + str(x) + ").JPG")
paginator = Paginator(data, 15)
page_number = request.GET.get('page')
page_obj = paginator.get_page(page_number)
if page_number is None:
no = 0
else:
no = int(page_number) * 15 - 15
context = {'data': page_obj, 'no': no, 'page': page_number}
return render(request, 'dataImage.html', context)
def load_data(result):
data = []
n, m = result.shape #活的图片大小
imap = np.zeros((m, n))
black = 0
for i in range(m):
for j in range(n): #将每个像素点RGB颜色处理到0-1范围内并存放data
point = result[j][i]
if not point == black:
x = result[j][i]
#y= result[j][i][1]
#z= result[j][i][2]
imap[i][j] = 1
data.append([x / 255.0])
return data, m, n, imap #以矩阵型式返回data，图片大小
def collectImages_for_dataset1(letter, freq):
speakers = ['A', 'B', 'C', 'D', 'E']
path = 'dataset1/'
data = []
for s in speakers:
image_path = path + s + '/' + letter + '/'
names = glob.glob(image_path + "*.*")
for i in range(0, len(names), freq):
name = names[i]
image = cv2.imread(name)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = convert_size_for_dataset1(image)
#plt.imshow(clear_background_rough(image))
data.append(image)
return data
def searchForImage(
self,
folderPath,
fileName,
functionOfSImilarity,
scale=1.0,
):
queryImage = LabImage(io.imread(folderPath + fileName), fileName)
queryImage.featuresValues = self.featuresExtractor(queryImage.img)
orderedIDs = [i for i in range(0, len(self.images))]
similarities = [
functionOfSImilarity(queryImage.featuresValues,
image.featuresValues) for image in self.images
]
tmp = list(zip(orderedIDs, similarities))
tmp.sort(key=lambda x: x[1], reverse=True)
orderedIDs, similarities = zip(*tmp)
### DISPLAY QUERY
figure(figsize=(5.0 * scale, 5.0 * scale), dpi=80)
subplot(1, 1, 1)
plt.imshow(queryImage.img)
plt.show()
print(queryImage.ID)
#features
data = []
for j in range(0, len(self.featuresNames)):
if isinstance(queryImage.featuresValues, np.ndarray):
fV = []
fV.append("- " + self.featuresNames[j])
fV.append("NaN")
data.append(fV)
else:
fV = []
fV.append("- " + self.featuresNames[j])
fV.append("{:10.2f}".format(queryImage.featuresValues[j]))
data.append(fV)
col_width = max(len(word) for row in data
for word in row) + 5 # PADDING
for row in data:
print("".join(word.ljust(col_width) for word in row))
### DISPLAY RESULTS
self.display(scale=scale,
orderedIDs=orderedIDs,
similarities=similarities)
def get_text(text, filename, size):
face = Face(filename)
face.set_char_size(POINT_SIZE(size))
pen = FT_Vector(0,0)
flags = FT_LOAD_RENDER
xmin,xmax,ymin,ymax = get_text_extents(text, filename, size)
L = np.zeros((ymax-ymin, xmax-xmin),dtype=np.ubyte)
previous = 0
pen.x, pen.y = 0, 0
for c in text:
import sys
sys.stdout.flush()
face.load_char(c, flags)
kerning = face.get_kerning(previous, c)
previous = c
bitmap = face.glyph.bitmap
pitch = face.glyph.bitmap.pitch
width = face.glyph.bitmap.width
rows = face.glyph.bitmap.rows
top = face.glyph.bitmap_top
left = face.glyph.bitmap_left
pen.x += kerning.x
x = (pen.x >> 6) - xmin + left
y = (pen.y >> 6) - ymin - (rows - top)
data = []
d = bitmap.buffer[:]
for j in range(rows):
data.append(d[j*pitch:j*pitch+width])
data = list(itertools.chain(data))
if len(data):
Z = np.array(data,dtype=np.ubyte).reshape(rows, width)
L[y:y+rows,x:x+width] |= Z[::-1,::1]
pen.x += face.glyph.advance.x
pen.y += face.glyph.advance.y
# Create a 10 pixel border for the image
L = np.flipud(L)
rows, cols = L.shape
row = 10*[[0]*cols]
L = np.vstack((row, L, row))
rows, cols = L.shape
col = np.array([[0]*10]*rows)
# print col.shape
L = np.hstack((col, L, col))
return L
def trainData():
print('training Started...')
data = []
labels = []
count = 0
dataFile = open('data.txt', 'w')
labelFile = open('labels.txt', 'w')
for car in vehicleList[:40000]:
# extract the make of the car
make = car.get('carType')
imagePath = car.get('imagePath')
# load the image, convert it to grayscale, and detect edges
image = cv2.imread(imagePath)
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
edged = imutils.auto_canny(gray)
# find contours in the edge map, keeping only the largest one which
# is presmumed to be the car logo
cnts = cv2.findContours(edged.copy(), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if imutils.is_cv2(cnts) else cnts[1]
c = max(cnts, key=cv2.contourArea)
# extract the logo of the car and resize it to a canonical width
# and height
(x, y, w, h) = cv2.boundingRect(c)
logo = gray[y:y + h, x:x + w]
logo = cv2.resize(logo, (200, 100))
# extract Histogram of Oriented Gradients from the logo
H = feature.hog(logo, orientations=9, pixels_per_cell=(10, 10),
cells_per_block=(2, 2), transform_sqrt=True, block_norm="L1")
# update the data and labels
data.append(H)
dataFile.write(str(H))
dataFile.write("\n")
labels.append(make)
labelFile.write(str(make))
labelFile.write("\n")
count += 1
return data, labels
def load_all_image_by_dir(path, conf):
start = t.time()
data = []
label = []
file_list = os.listdir(path)
debug_count = 0
for file in file_list:
# debug_count+=1
# if debug_count>100: break #调试用，省的加载那么长时间图片
if (file.find(".jpg") == -1 and file.find(".jpeg") == -1
and file.find(".png") == -1 and file.find(".bmp") == -1
and file.find(".gif") == -1):
continue
file_name = file.split(".")[0]
img_data = cv2.imread(path + file)
one_img = preprocess_image(img_data, file_name, conf)
if (one_img is None):
logger.debug("处理图像失败%s，处理下一张", file)
continue
try:
v = label_process.label2vector(file_name, conf)
label.append(v)
except Exception as e:
import traceback
traceback.print_exc()
logger.error("忽略此文件%s,原因：%r", file_name, str(e))
continue
data.append(one_img)
#把数组堆叠到一起形成一个[20000,100,36,1]的张量
# print(len(data))
image_data = np.vstack(data)
label_data = np.vstack(label)
logger.info("images data loaded:%r", image_data.shape)
end = t.time()
logger.info("加载图像使用了%d秒...", (end - start))
return image_data, label_data
def gen_dataset(n=df.shape[0]):
data = []
labels = []
for i in range(1, n):
crop, success = getcrop(i)
if (success):
data.append(crop)
labels.append(df[2][i])
else:
if (enable_error_output):
print("[WARNING] Template matching has failed for image: " +
str(i))
print_percentage((i * 100 / (n - 1)), "Fetched " + str(i) + " images:")
print_percentage(100, "Fetched " + str(n - 1) + " images:")
print("")
print("Finished!")
return data, labels
def rr_img(img):
'''旋转和去除离群点'''
debug = False
img = loose.utils.get_rotate_img(img)
if debug:
cv.imshow('src_img', img)
# 计算图片吊弦所在的x坐标
if len(img) > 0:
h, w = img.shape
center = []
for j in range(h):
row = []
for i in range(w):
if img[j][i] == 255:
row.append(i)
if len(row) > 0:
center.append(np.mean(row))
center_x = np.mean(center).astype(np.int)
# 截取吊弦两侧blank像素的范围
blank = 20
for j in range(h):
for i in range(0, center_x - blank):
img[j][i] = 0
for i in range(center_x + blank, w):
img[j][i] = 0
if debug: cv.imshow('blank', img)
# # 去除离散点
result = img > 1
result = morphology.remove_small_objects(result, min_size=20, connectivity=1)
# bool 图片到常规图片
data = []
nres = np.zeros(result.shape, dtype=np.uint8)
for i in range(result.shape[0]):
for j in range(result.shape[1]):
if result[i][j] == False:
nres[i][j] = 0
else:
nres[i][j] = 255
data.append([i, j])
if debug: cv.imshow('rm oulier', nres)
return nres
def get_arrowheads(img, lines):
img = rgb2grey(img)
img = skimage.img_as_ubyte(img)
fig, ax = plt.subplots()
fig.tight_layout()
data = []
relevant = []
id_counter = 0
# Enden alle Geraden betrachten
for l in lines:
start, end = l[1], l[2]
dst_start, dst_end, index_start, index_end = has_arrowhead(
l[0], start, end, img, id_counter)
if (dst_start != -1) and (dst_end != -1):
data.append(
(dst_start, dst_end, start, end, index_start, index_end, l[0]))
id_counter = index_end
# Pfeilspitzenpaare sortieren nach Gesamtdistanz zur Template Pfeilspitze
data = [
i for i in data
if (i[0] > ARROW_THRESHOLD) and (i[1] > ARROW_THRESHOLD)
]
#data = sorted(data, key=lambda x: x[0] + x[1], reverse=True)
#data = data[0:10]
# Auswahl auf Pfeilspitzenpaare mit geringstem Abstand zur Template Pfeilspitze beschränken
for d in data:
# Bounding Boxen der Pfeilspitzen einzeichnen
draw_bounding_box(ax, img, d[2], d[3], d[4], d[5], d[6])
relevant.append((d[2][1], d[2][0]))
relevant.append((d[3][1], d[3][0]))
#print(data)
print(len(data))
ax.imshow(img, interpolation='bicubic', cmap=plt.cm.gray)
#plt.show()
return relevant
def load_filenames_process_and_save_results(filename, flag_use_skimage_version):
# load samples filenames
with open("data\\samples_filenames\\" + filename + "_filenames.pickle", "rb") as f:
samples_filenames = pickle.load(f)
data = []
for sample_filename in samples_filenames:
single_data = HOG_function(sample_filename, flag_use_skimage_version)
data.append(single_data)
print(len(data))
if flag_use_skimage_version:
output_filename_modifier = ""
else:
output_filename_modifier = "_modified"
with open("data\\" + filename + output_filename_modifier + ".pickle", "wb") as f:
pickle.dump(data, f)
return data
def line_fit(frame):
'''
获取距离变换，水平投影和垂直投影后的骨架图
@param frame: cv.img color
@return:
nres: 骨架图，gray
data: 图像中前景坐标
'''
if len(frame.shape)==2:
frame = cv.cvtColor(frame,cv.COLOR_GRAY2BGR)
# 二值化图像
gray = cv.cvtColor(frame, cv.COLOR_BGR2GRAY)
gray = cv.GaussianBlur(gray,(5,5),0)
ret, binary = cv.threshold(gray, 3, 255, cv.THRESH_BINARY)
# 距离变换，投影
result,xpts,ypts = projection(binary)
# 每一行前景如果像素值多于1，则该行置为0
cnt = np.sum(result,-1)
idx = np.where(cnt>255)
result[idx, :] = 0
space = 15
result = result[space:-space,space:-space]
## 去除离散点
result = result >1
result = morphology.remove_small_objects(result, min_size=2, connectivity=1)
## 布尔图还原到灰度图
data = []
nres = np.zeros(result.shape,dtype=np.uint8)
for i in range(result.shape[0]):
for j in range(result.shape[1]):
if result[i][j] == False:
nres[i][j]=0
else:
nres[i][j]=255
data.append([i,j])
return nres,data
croppedImage = currentImage[(sizeOfImage[0]-sizeOfNewImage[0])/2:(sizeOfImage[0]+sizeOfNewImage[0])/2,(sizeOfImage[1]-sizeOfNewImage[1])/2:(sizeOfImage[1]+sizeOfNewImage[1])/2]
flippedImageH = flipImageHorizontally(croppedImage)
flippedImageV = flipImageVertically(croppedImage)
currentFolder.append(resizedImage)
currentFolderLabels.append(index)
currentFolder.append(croppedImage)
currentFolderLabels.append(index)
currentFolder.append(flippedImageH)
currentFolderLabels.append(index)
currentFolder.append(flippedImageV)
currentFolderLabels.append(index)
if laplaceImages:
currentLaplaceImage = scImage.laplace(currentImage)
currentFolder.append(currentLaplaceImage)
currentFolderLabels.append(index)
data.append(currentFolder)
labels.append(currentFolderLabels)
index = index + 1
print("DATA READING DONE")
index = 0
for i in labels:
labels[index] = np_utils.to_categorical(i, 15)
index = index + 1
print("LABELS WERE CHANGED")
#MAKE TRAINING AND TEST
Xtraining = []
Ytraining = []
#Xtest = []
#Ytest = []
index = 0
import matplotlib.pyplot as plt
from skimage.feature import hog
from skimage import data, color, exposure
from scipy import misc
from PIL import Image
from resizeimage import resizeimage
from sklearn import svm
from sklearn.cross_validation import KFold
f = open('/Users/yanan/caffe/data/vehicle/train.txt')
data = []
label = []
feature = []
for line in f:
data.append(line.split(' ')[0])
label.append(line.split(' ')[1].strip())
for d in data:
image = misc.imread(d)
fd, hog_image = hog(image, orientations=8, pixels_per_cell=(4, 4),
cells_per_block=(1, 1), visualise=True)
feature.append(fd)
kf = KFold(len(data), n_folds=10, shuffle=True)
for train_index, test_index in kf:
tmp_feature = [feature[x] for x in train_index]
tmp_label = [label[x] for x in train_index]
test_feature = [feature[x] for x in test_index]
test_label = [label[x] for x in test_index]
clf = svm.SVC()
def proj3_22b():
all_channels = []
for channel in range(3):
bg = plt.imread("data/trumpsou.jpeg") / 255
#bg = exposure.equalize_hist(bg)
bg = np.clip(bg + .2 , 0, 1)
bg = bg[...,channel]
source = ndimage.rotate(plt.imread("data/kavi.jpg"), -90) / 255
source = exposure.equalize_hist(source)
source = np.clip(source + .2, 0, 1)
source = source[...,channel]
#plt.imshow(bg, cmap="gray")
#plt.show()
#plt.imshow(source, cmap="gray")
#plt.show()
# How to trim the input image
sx1 = 0 # sx1 = 20
sx2 = source.shape[0] # sx2 = 450
sy1 = 0 # sy1 = 50
sy2 = source.shape[1] # sy2 = 360
# How much to scale the input image down
#K = .052
K = .04
# Top left corner on the source image in bg image frame
#full_bgx1 = 100
full_bgx1 = 135
#full_bgy1 = 450
full_bgy1 = 130
# Corresponding bottom right corner
full_bgx2 = int(full_bgx1 + (sx2 - sx1) * K)
full_bgy2 = int(full_bgy1 + (sy2 - sy1) * K)
bgx1 = full_bgx1 + 10
bgy1 = full_bgy1 + 20
bgx2 = full_bgx2 - 5
bgy2 = full_bgy2 - 2
#penguin = source[sx1:sx2,sy1:sy2]
penguin = source
penguin = imresize(penguin, (full_bgx2 - full_bgx1, full_bgy2 - full_bgy1)) / 255
# print("AHHHHH", full_bgx1, full_bgx2, full_bgy1, full_bgy2)
# print("SOURCE: ", penguin.shape)
# print("BG: ", bg.shape)
bg_with_full_source = bg.copy()
bg_with_full_source[full_bgx1:full_bgx2, full_bgy1:full_bgy2] = penguin
plt.imshow(bg_with_full_source, cmap="gray")
plt.show()
scipy.misc.imsave("output/final_original_kavitrump.jpg", bg_with_full_source)
#exit(0)
#plt.imshow(bg, cmap="gray")
#plt.show()
#exit(0)
#plt.imshow(penguin)
#plt.show()
#bg[bgx1:bgx2, bgy1:bgy2] = penguin
#plt.imshow(bg, cmap="gray")
#plt.show()
# exit(0)
row_num = 0
row = []
col = []
data = []
b = []
# where we're traversing
start = (bgx1, bgy1)
end = (bgx2, bgy2)
source_shape = (bgx2 - bgx1, bgy2 - bgy1)
for i in range(bgx1, bgx2):
for j in range(bgy1, bgy2):
# CHECK DOWN (i + 1)
if i < bgx2 - 1:
idx1 = get_idx(i, j, start, end)
idx2 = get_idx(i + 1, j, start, end)
row.append(row_num)
col.append(idx1)
data.append(1)
row.append(row_num)
col.append(idx2)
data.append(-1)
b.append(bg_with_full_source[i,j] - bg_with_full_source[i+1, j])
row_num += 1
# if its on the very last row, get the other i+1 equation
elif i == bgx2 - 1:
idx1 = get_idx(i, j, start, end)
row.append(row_num)
col.append(idx1)
data.append(1)
b.append(bg_with_full_source[i,j] - bg_with_full_source[i+1, j] + bg[i+1,j])
row_num += 1
# CHECK RIGHT (j + 1)
if j < bgy2 - 1:
idx1 = get_idx(i, j, start, end)
idx2 = get_idx(i, j + 1, start, end)
row.append(row_num)
col.append(idx1)
data.append(1)
row.append(row_num)
col.append(idx2)
data.append(-1)
b.append(bg_with_full_source[i,j] - bg_with_full_source[i, j+1])
row_num += 1
elif j == bgy2 - 1:
idx1 = get_idx(i, j, start, end)
row.append(row_num)
col.append(idx1)
data.append(1)
b.append(bg_with_full_source[i,j] - bg_with_full_source[i, j+1] + bg[i,j+1])
row_num += 1
# CHECK UP (i - 1)
if i > bgx1:
idx1 = get_idx(i, j, start, end)
idx2 = get_idx(i - 1, j, start, end)
row.append(row_num)
col.append(idx1)
data.append(1)
row.append(row_num)
col.append(idx2)
data.append(-1)
b.append(bg_with_full_source[i,j] - bg_with_full_source[i-1, j])
row_num += 1
# if its on the very last row, get the other i+1 equation
elif i == bgx1:
idx1 = get_idx(i, j, start, end)
row.append(row_num)
col.append(idx1)
data.append(1)
b.append(bg_with_full_source[i,j] - bg_with_full_source[i-1, j] + bg[i-1,j])
row_num += 1
# CHECK UP (j - 1)
if j > bgy1:
idx1 = get_idx(i, j, start, end)
idx2 = get_idx(i, j-1, start, end)
row.append(row_num)
col.append(idx1)
data.append(1)
row.append(row_num)
col.append(idx2)
data.append(-1)
b.append(bg_with_full_source[i,j] - bg_with_full_source[i, j-1])
row_num += 1
# if its on the very last row, get the other i+1 equation
elif j == bgy1:
idx1 = get_idx(i, j, start, end)
row.append(row_num)
col.append(idx1)
data.append(1)
b.append(bg_with_full_source[i,j] - bg_with_full_source[i, j-1] + bg[i,j-1])
row_num += 1
row, col, data = np.array(row), np.array(col), np.array(data)
A = csr_matrix((data, (row, col)))
b = np.array([b]).T
print("Constructed A matrix of shape: ", A.shape)
print("Constructed b matrix of shape: ", b.shape)
x = la.lsqr(A, b)
y = x[0]
print("Pixel values found: ", y.shape)
final_source = np.reshape(y, source_shape)
#plt.imshow(final_source, cmap="gray")
#plt.show()
bg_channel = bg.copy()
bg_channel[bgx1:bgx2, bgy1:bgy2] = final_source
#plt.imshow(bg_channel, cmap="gray")
#plt.show()
all_channels.append(bg_channel)
final_im = np.empty((all_channels[0].shape[0], all_channels[0].shape[1], 3))
for channel in range(3):
final_im[:,:,channel] = np.clip(all_channels[channel], 0, 1)
plt.imshow(final_im)
plt.show()
scipy.misc.imsave("output/final_kavitrump.jpg", final_im)
def train_ai():
data = []
classID = []
features = []
features_temp_array = []
'''
#SIMPLECV
#bows
feature_extractors = []
extractor_names = []
# Training data set paths for classification(suppervised learnning)
image_dirs = ['../coin_images/jheads/',
'../coin_images/jtails/',
'../coin_images/oheads/',
'../coin_images/otails/',
]
# Different class labels for multi class classification
class_names = ['jhead','jtail','ohead', 'otail']
#preprocess all training images
for directory in image_dirs:
for filename in glob.glob(directory + '/*.jpg'):
print "Processing:", filename
img = cv2.imread(filename)
temp_img = preprocess_houghlines (img, 100)
temp_str = filename.rsplit('/')
temp_str = temp_str[len(temp_str)-1]
temp_str = directory + '/temp/' + temp_str
print temp_str
cv2.imwrite(temp_str, temp_img)
#raw_input('press enter to continue : ')
#sys.exit(-1)
#build array of directories for bow
#image_dirs2 = []
#for directory in image_dirs:
# image_dirs2.append(directory + '/temp/')
#print image_dirs2
# Different class labels for multi class classification
extractor_name = 'hue'
if extractor_name == 'bow':
feature_extractor = BOFFeatureExtractor() # feature extrator for bag of words methodology
feature_extractor.generate(image_dirs,imgs_per_dir=40) # code book generation
elif extractor_name == 'hue':
feature_extractor = HueHistogramFeatureExtractor()
elif extractor_name == 'morphology':
feature_extractor = MorphologyFeatureExtractor()
elif extractor_name == 'haar':
feature_extractor = HaarLikeFeatureExtractor()
elif extractor_name == 'edge':
feature_extractor = EdgeHistogramFeatureExtractor()
image_dirs2 = image_dirs
#bow_features = BOFFeatureExtractor()
#bow_features.generate(image_dirs2,imgs_per_dir=40, verbose=True) # code book generation
#bow_features.generate(image_dirs2,imgs_per_dir=200,numcodes=256,sz=(11,11),img_layout=(16,16),padding=4 )
#bow_features.save('codebook.png','bow.txt')
#print "extractor_names:", extractor_names, feature_extractors
# initializing classifier with appropriate feature extractors list
#print type(bow_features), bow_features, bow_features.getFieldNames(), bow_features.getNumFields()
#raw_input('bow saved...Enter : ')
#bow_features = None
#bow_features = BOFFeatureExtractor()
#print type(bow_features), bow_features, bow_features.getFieldNames(), bow_features.getNumFields()
#bow_features.load('bow.txt')
#print type(bow_features), bow_features, bow_features.getFieldNames(), bow_features.getNumFields()
feature_extractors.append(feature_extractor)
#raw_input('bow loaded Enter : ')
#extractor_names.append(extractor_name)
classifier_name = 'naive'
if classifier_name == 'naive':
classifier = NaiveBayesClassifier(feature_extractors)
elif classifier_name == 'svm':
classifier = SVMClassifier(feature_extractors)
elif classifier_name == 'knn':
classifier = KNNClassifier(feature_extractors, 2)
elif classifier_name == 'tree':
classifier = TreeClassifier(feature_extractors)
# train the classifier to generate hypothesis function for classification
#print "image_dirs:", image_dirs, class_names
classifier.train(image_dirs2,class_names,disp=None,savedata='features.txt',verbose=True)
print 'classifier:', type(classifier), classifier
raw_input('press enter to continue :')
#pickle.dump( classifier, open( "coinvision_ai_model2.mdl", "wb" ),2 )
#classifier.save('coinvision_ai_model.mdl')
print 'classifier:', type(classifier), classifier
#classifier = NaiveBayesClassifier.load('coinvision_ai_model.mdl')
#raw_input('press enter to continue : let me try loading bow file')
#classifier2 = NaiveBayesClassifier.load('coinvision_ai_model.mdl')
#classifier2.setFeatureExtractors(feature_extractors)
#print 'classifier2:', type(classifier2), classifier2
#classifier.load("coinvision_ai_model.mdl")
#classifier2.load('coinvision_ai_model.mdl')
#print 'classifier:', type(classifier2), classifier2
raw_input('press enter to continue : ')
print 'testing ai:'
test_images_path = "../coin_images/unclassified"
extension = "*.jpg"
if not test_images_path:
path = os.getcwd() #get the current directory
else:
path = test_images_path
directory = os.path.join(path, extension)
files = glob.glob(directory)
count = 0 # counting the total number of training images
error = 0 # conuting the total number of misclassification by the trained classifier
for image_file in files:
new_image = Image(image_file)
category = classifier.classify(new_image)
print "image_file:", image_file + " classified as: " + category
if image_file[-9] == 't':
if category == 'jhead' or category == 'ohead':
print "INCORRECT CLASSIFICATION"
error += 1
if image_file[-9] == 'h':
if category == 'jtail' or category == 'otail':
print "INCORRECT CLASSIFICATION"
error += 1
count += 1
# reporting the results
print ' * classifier : ', classifier
print ' * extractors :', extractor_names
print ' *', error, 'errors out of', count
raw_input('edned press enter to continue : ')
return
'''
#try:
data_filename = 'coinvision_feature_data.csv'
print 'reading features and classID: ', data_filename
f_handle = open(data_filename, 'r')
reader = csv.reader(f_handle)
#read data from file into arrays
for row in reader:
data.append(row)
for row in range(0, len(data)):
#print features[row][1]
classID.append(int(data[row][0]))
features_temp_array.append(data[row][1].split(" "))
#removes ending element which is a space
for x in range(len(features_temp_array)):
features_temp_array[x].pop()
features_temp_array[x].pop(0)
#convert all strings in array to numbers
temp_array = []
for x in range(len(features_temp_array)):
temp_array = [ float(s) for s in features_temp_array[x] ]
features.append(temp_array)
#make numpy arrays
features = np.asarray(features)
#print classID, features
learner = milk.defaultclassifier(mode='really-slow')
model = learner.train(features, classID)
pickle.dump( model, open( "coinvision_ai_model.mdl", "wb" ) )
#except:
print "could not retrain.. bad file"
from sklearn import svm
model = svm.SVC(gamma=0.001, C=100.)
model.fit(features, classID)
pickle.dump( model, open( "coinvision_ai_model_svc.mdl", "wb" ) )
from sklearn.neighbors import KNeighborsClassifier
neigh = KNeighborsClassifier(n_neighbors=3)
neigh.fit(features, classID)
pickle.dump( neigh, open( "coinvision_ai_model_knn.mdl", "wb" ) )
from sklearn.svm import LinearSVC
clf = LinearSVC()
clf = clf.fit(features, classID)
pickle.dump( clf, open( "coinvision_ai_model_lr.mdl", "wb" ) )
from sklearn.linear_model import LogisticRegression
clf2 = LogisticRegression().fit(features, classID)
pickle.dump( clf2 , open( "coinvision_ai_model_clf2.mdl", "wb" ) )
return
